Adversarial learning of image augmentation policies for object detection

Abstract

Recent deep learning architectures have achieved impressive results on the task of Object Detection. However, such techniques are known to require huge amounts of labeled data in order to properly generalize, which is harder to acquire than traditional image classification data. To improve this situation, the use of image augmentation techniques became standard practice on the literature. This traditional forms of augmentation consist mostly of fixed sequences of handcrafted image operations with random parameters, which already helps counteract the lack of labeled data on generic object detection, although not being enough in more specific, low data scenarios. In this perspective, a few works already apply image generation techniques to artificially create novel training samples, but to the best of our knowledge, none of them addresses object detection yet.Parallel to this, recently proposed generative models, based on the GAN framework, have achieved impressive results in terms of image realism for some classes of objects, and works on image augmentation already employ them for classification tasks.The goal of the proposed internship is to investigate the application of recently proposed generative adversarial architectures for the automatic generation of labeled samples, in a way that improves the performance of deep learning based object detection models. (AU)